X-Outline of a Theory of Statistical Estimation

While visiting Warwick last week, Jean-Michel Marin pointed out and forwarded me this remarkable paper of Jerzy Neyman, published in 1937, and presented to the Royal Society by Harold Jeffreys.

“Leaving apart on one side the practical difficulty of achieving randomness and the meaning of this word when applied to actual experiments…”

“It may be useful to point out that although we are frequently witnessing controversies in which authors try to defend one or another system of the theory of probability as the only legitimate, I am of the opinion that several such theories may be and actually are legitimate, in spite of their occasionally contradicting one another. Each of these theories is based on some system of postulates, and so long as the postulates forming one particular system do not contradict each other and are sufficient to construct a theory, this is as legitimate as any other. “

This paper is fairly long in part because Neyman starts by setting Kolmogorov’s axioms of probability. This is of historical interest but also needed for Neyman to oppose his notion of probability to Jeffreys’ (which is the same from a formal perspective, I believe!). He actually spends a fair chunk on explaining why constants cannot have anything but trivial probability measures. Getting ready to state that an a priori distribution has no meaning (p.343) and that in the rare cases it does it is mostly unknown. While reading the paper, I thought that the distinction was more in terms of frequentist or conditional properties of the estimators, Neyman’s arguments paving the way to his definition of a confidence interval. Assuming repeatability of the experiment under the same conditions and therefore same parameter value (p.344).

“The advantage of the unbiassed [sic] estimates and the justification of their use lies in the fact that in cases frequently met the probability of their differing very much from the estimated parameters is small.”

“…the maximum likelihood estimates appear to be what could be called the best “almost unbiassed [sic]” estimates.”

It is also quite interesting to read that the principle for insisting on unbiasedness is one of producing small errors, because this is not that often the case, as shown by the complete class theorems of Wald (ten years later). And that maximum likelihood is somewhat relegated to a secondary rank, almost unbiased being understood as consistent. A most amusing part of the paper is when Neyman inverts the credible set into a confidence set, that is, turning what is random in a constant and vice-versa. With a justification that the credible interval has zero or one coverage, while the confidence interval has a long-run validity of returning the correct rate of success. What is equally amusing is that the boundaries of a credible interval turn into functions of the sample, hence could be evaluated on a frequentist basis, as done later by Dennis Lindley and others like Welch and Peers, but that Neyman fails to see this and turn the bounds into hard values. For a given sample.

“This, however, is not always the case, and in general there are two or more systems of confidence intervals possible corresponding to the same confidence coefficient α, such that for certain sample points, E’, the intervals in one system are shorter than those in the other, while for some other sample points, E”, the reverse is true.”

The resulting construction of a confidence interval is then awfully convoluted when compared with the derivation of an HPD region, going through regions of acceptance that are the dual of a confidence interval (in the sampling space), while apparently [from my hasty read] missing a rule to order them. And rejecting the notion of a confidence interval being possibly empty, which, while being of practical interest, clashes with its frequentist backup.

exact, unbiased, what else?!

Last week, Matias Quiroz, Mattias Villani, and Robert Kohn arXived a paper on exact subsampling MCMC, a paper that contributes to the current literature on approximating MCMC samplers for large datasets, in connection with an earlier paper of Quiroz et al. discussed here last week.

The “exact” in the title is to be understood in the Russian roulette sense. By using Rhee and Glynn debiaising device, the authors achieve an unbiased estimator of the likelihood as in Bardenet et al. (2015). The central tool for the derivation of an unbiased and positive estimator is to find a control variate for each component of the log likelihood that is good enough for the difference between the component and the control to be lower bounded. By the constant a in the screen capture above. When the individual terms d in the product are iid unbiased estimates of the log likelihood difference. And q is the sum of the control variates. Or maybe more accurately of the cheap substitutes to the exact log likelihood components. Thus still of complexity O(n), which makes the application to tall data more difficult to contemplate.

The $64 question is obviously how to produce cheap and efficient control variates that kill the curse of the tall data. (It still irks to resort to this term of control variate, really!) Section 3.2 in the paper suggests clustering the data and building an approximation for each cluster, which seems to imply manipulating the whole dataset at this early stage. At a cost of O(Knd). Furthermore, because finding a correct lower bound a is close to impossible in practice, the authors use a “soft lower bound”, meaning that it is only an approximation and thus that (3.4) above can get negative from time to time, which cancels the validation of the method as a pseudo-marginal approach. The resolution of this difficulty is to resort to the same proxy as in the Russian roulette paper, replacing the unbiased estimator with its absolute value, an answer I already discussed for the Russian roulette paper. An additional step is proposed by Quiroz et al., namely correlating the random numbers between numerator and denominator in their final importance sampling estimator, via a Gaussian copula as in Deligiannidis et al.

This paper made me wonder (idly wonder, mind!) anew how to get rid of the vexing unbiasedness requirement. From a statistical and especially from a Bayesian perspective, unbiasedness is a second order property that cannot be achieved for most transforms of the parameter θ. And that does not keep under reparameterisation. It is thus vexing and perplexing that unbiased is so central to the validation of our Monte Carlo technique and that any divergence from this canon leaves us wandering blindly with no guarantee of ever reaching the target of the simulation experiment…

MCMskv #3 [town with a view]

Third day at MCMskv, where I took advantage of the gap left by the elimination of the Tweedie Race [second time in a row!] to complete and submit our mixture paper. Despite the nice weather. The rest of the day was quite busy with David Dunson giving a plenary talk on various approaches to approximate MCMC solutions, with a broad overview of the potential methods and of the need for better solutions. (On a personal basis, great line from David: “five minutes or four minutes?”. It almost beat David’s question on the previous day, about the weight of a finch that sounded suspiciously close to the question about the air-speed velocity of an unladen swallow. I was quite surprised the speaker did not reply with the Arthurian “An African or an European finch?”) In particular, I appreciated the notion that some problems were calling for a reduction in the number of parameters, rather than the number of observations. At which point I wrote down “multiscale approximations required” in my black pad,  a requirement David made a few minutes later. (The talk conditions were also much better than during Michael’s talk, in that the man standing between the screen and myself was David rather than the cameraman! Joke apart, it did not really prevent me from reading them, except for most of the jokes in small prints!)

The first session of the morning involved a talk by Marc Suchard, who used continued fractions to find a closed form likelihood for the SIR epidemiology model (I love continued fractions!), and a talk by Donatello Telesca who studied non-local priors to build a regression tree. While I am somewhat skeptical about non-local testing priors, I found this approach to the construction of a tree quite interesting! In the afternoon, I obviously went to the intractable likelihood session, with talks by Chris Oates on a control variate method for doubly intractable models, Brenda Vo on mixing sequential ABC with Bayesian bootstrap, and Gael Martin on our consistency paper. I was not aware of the Bayesian bootstrap proposal and need to read through the paper, as I fail to see the appeal of the bootstrap part! I later attended a session on exact Monte Carlo methods that was pleasantly homogeneous. With talks by Paul Jenkins (Warwick) on the exact simulation of the Wright-Fisher diffusion, Anthony Lee (Warwick) on designing perfect samplers for chains with atoms, Chang-han Rhee and Sebastian Vollmer on extensions of the Glynn-Rhee debiasing technique I previously discussed on the blog. (Once again, I regretted having to make a choice between the parallel sessions!)

The poster session (after a quick home-made pasta dish with an exceptional Valpolicella!) was almost universally great and with just the right number of posters to go around all of them in the allotted time. With in particular the Breaking News! posters of Giacomo Zanella (Warwick), Beka Steorts and Alexander Terenin. A high quality session that made me regret not touring the previous one due to my own poster presentation.

Bayes and unbiased

Siamak Noorbaloochi and Glen Meeden have arXived a note on some mathematical connections between Bayes estimators and unbiased estimators.  To start with, I never completely understood the reasons for [desperately] seeking unbiasedness in estimators, given that most transforms of a parameter θ do not allow for unbiased estimators when based on a finite sample from a parametric family with such parameter θ (Lehmann, 1983). (This is not even for a Bayesian reason!) The current paper first seems to use unbiasedness in a generalised sense introduced earlier by the authors in 1983, which is a less intuitive notion since it depends on loss and prior, still without being guaranteed to exist since it involves an infimum over a non-compact space. However, for the squared error loss adopted in this paper, it seems to reduce to the standard notion.

A first mathematical result therein is that the Bayes [i.e., posterior mean] and unbiasedness [i.e., sample mean] operators are adjoint in a Hilbert sense. But this does not seem much more than a consequence of Fubini’s theorem. The authors then proceeds to expose the central (decomposition) result of the paper, namely that every estimable function γ(θ) can be orthogonally decomposed into a function with an unbiased estimator plus a function whose Bayes estimator is zero and that conversely every square-integrable estimator can be orthogonally decomposed into a Bayes estimator (of something) plus an unbiased estimator of zero. This is a neat duality result, whose consequences I however fail to see because the Bayes estimator is estimating something else. And, somewhere, somehow, I have some trouble with the notion of a function α whose Bayes estimator [i.e., posterior mean] is zero for all values of the sample, esp. outside problems with finite observation and finite parameter spaces. For instance, if the sampling distribution is within an exponential family, the above property means that the Laplace transform of this function α is uniformly zero, hence that the function itself is uniformly zero.

SMC 2015

Nicolas Chopin ran a workshop at ENSAE on sequential Monte Carlo the past three days and it was a good opportunity to get a much needed up-to-date on the current trends in the field. Especially given that the meeting was literally downstairs from my office at CREST. And given the top range of researchers presenting their current or past work (in the very amphitheatre where I attended my first statistics lectures, a few dozen years ago!). Since unforeseen events made me miss most of the central day, I will not comment on individual talks, some of which I had already heard in the recent past, but this was a high quality workshop, topped by a superb organisation. (I started wondering why there was no a single female speaker in the program and so few female participants in the audience, then realised this is a field with a massive gender imbalance, which is difficult to explain given the different situation in Bayesian statistics and even in Bayesian computation…)  Some key topics I gathered during the talks I could attend–apologies to the other speakers for missing their talk due to those unforeseen events–are unbiasedness, which sounds central to the SMC methods [at least those presented there] as opposed to MCMC algorithms, and local features, used in different ways like hierarchical decomposition, multiscale, parallelisation, local coupling, &tc., to improve convergence and efficiency…

an extension of nested sampling

I was reading [in the Paris métro] Hastings-Metropolis algorithm on Markov chains for small-probability estimation, arXived a few weeks ago by François Bachoc, Lionel Lenôtre, and Achref Bachouch, when I came upon their first algorithm that reminded me much of nested sampling: the following was proposed by Guyader et al. in 2011,

To approximate a tail probability P(H(X)>h),

• start from an iid sample of size N from the reference distribution;
• at each iteration m, select the point x with the smallest H(x)=ξ and replace it with a new point y simulated under the constraint H(y)≥ξ;
• stop when all points in the sample are such that H(X)>h;
• take

as the unbiased estimator of P(H(X)>h).

Hence, except for the stopping rule, this is the same implementation as nested sampling. Furthermore, Guyader et al. (2011) also take advantage of the bested sampling fact that, if direct simulation under the constraint H(y)≥ξ is infeasible, simulating via one single step of a Metropolis-Hastings algorithm is as valid as direct simulation. (I could not access the paper, but the reference list of Guyader et al. (2011) includes both original papers by John Skilling, so the connection must be made in the paper.) What I find most interesting in this algorithm is that it even achieves unbiasedness (even in the MCMC case!).

PMC for combinatoric spaces

I received this interesting [edited] email from Xiannian Fan at CUNY:

I am trying to use PMC to solve Bayesian network structure learning problem (which is in a combinatorial space, not continuous space).

In PMC, the proposal distributions q_i,t can be very flexible, even specific to each iteration and each instance. My problem occurs due to the combinatorial space.

For importance sampling, the requirement for proposal distribution, q, is:

support (p) ⊂ support (q)             (*)

For PMC, what is the support of the proposal distribution in iteration t? is it

support (p) ⊂ U support(q_i,t)    (**)

or does (*) apply to every q_i,t?

For continuous problem, this is not a big issue. We can use random walk of Normal distribution to do local move satisfying (*). But for combination search, local moving only result in finite states choice, just not satisfying (*). For example for a permutation (1,3,2,4), random swap has only choose(4,2)=6 neighbor states.

Fairly interesting question about population Monte Carlo (PMC), a sequential version of importance sampling we worked on with French colleagues in the early 2000’s.  (The name population Monte Carlo comes from Iba, 2000.)  While MCMC samplers do not have to cover the whole support of p at each iteration, it is much harder for importance samplers as their core justification is to provide an unbiased estimator to for all integrals of interest. Thus, when using the PMC estimate,

1/n ∑_i,t {p(x_i,t)/q_i,t(x_i,t)}h(q_i,t),  x_i,t~q_i,t(x)

this estimator is only unbiased when the supports of the q_i,t “s are all containing the support of p. The only other cases I can think of are

1. associating the q_i,t “s with a partition S_i,t of the support of p and using instead

   ∑_i,t {p(x_i,t)/q_i,t(x_i,t)}h(q_i,t), x_i,t~q_i,t(x)

2. resorting to AMIS under the assumption (**) and using instead

   1/n ∑_i,t {p(x_i,t)/∑_j,t q_j,t(x_i,t)}h(q_i,t), x_i,t~q_i,t(x)

but I am open to further suggestions!